DeepSeek trained V4 on Huawei chips rather than NVIDIA GPUs, signaling that China's AI can move toward chip independence. Released in June 2026, DeepSeek-V4-Pro is a 1.6-trillion-parameter Mixture-of-Experts model, the first major frontier-class Chinese model trained at scale on Huawei Ascend 950 chips. It signals that US chip export controls are not stopping China from training advanced models. ASAP summarizes the result from the primary source.

Trained without NVIDIA

The core of DeepSeek V4 is that the training chip changed. It is the first major frontier-class Chinese model trained at scale on Huawei Ascend 950 chips rather than NVIDIA hardware. The bigger news is not the model itself but where it was built.

V4 in numbers

DeepSeek-V4-Pro is a 1.6-trillion-parameter Mixture-of-Experts model. The total parameter count reaches 1.6 trillion, but only some experts activate at inference to raise efficiency. It effectively proves that training at this scale is possible on the Huawei platform.

Bypassing export controls

The training of V4 is direct evidence that US chip export controls are failing to achieve their intended effect. The US restricted exports of advanced GPUs to China in 2026, yet DeepSeek trained a 1.6-trillion-parameter model on domestic Huawei chips. Blocking hardware does not directly translate into blocking capability.

Still short of the US frontier

On performance, V4 is still behind the top US closed models, by a clear margin. A Council on Foreign Relations analysis rated V4 as the best available open-source option, but found it not competitive with US frontier closed models. Chip independence and top performance are different problems.

What it means — a sign of chip independence

DeepSeek V4 shows that China has secured a path to train large models without NVIDIA. The performance gap remains, but the fact that the Huawei platform can serve large-scale training is itself a structural shift. The variable in the AI race is moving from the model to the chip supply chain.

Wrap-up

The real news of DeepSeek V4 is not performance but the chip. A 1.6-trillion-parameter MoE model was trained on Huawei Ascend 950 without NVIDIA, and export controls did not stop it. It still trails the US frontier, but it signals that China's path to chip independence is open.

Source: DeepSeek "V4-Pro" release (June 2026; 1.6-trillion-parameter MoE, trained on Huawei Ascend 950) and Council on Foreign Relations analysis, summarized by ASAP.